pocketfft.cc 34 KB
Newer Older
Martin Reinecke's avatar
Martin Reinecke committed
1
2
3
4
5
6
#include <cmath>
#include <cstring>
#include <algorithm>
#include <iostream>
#include <memory>
#include <vector>
Martin Reinecke's avatar
Martin Reinecke committed
7
8
#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>
Martin Reinecke's avatar
Martin Reinecke committed
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37

#ifdef __GNUC__
#define NOINLINE __attribute__((noinline))
#define restrict __restrict__
#else
#define NOINLINE
#define restrict
#endif

using namespace std;

namespace {

template<typename T> struct arr
  {
  private:
    T *p;
    size_t sz;

    static T *ralloc(size_t num)
      { return (T *)aligned_alloc(64,num*sizeof(T)); }
    static void dealloc(T *ptr)
      { free(ptr); }

  public:
    arr() : p(0), sz(0) {}
    arr(size_t n) : p(ralloc(n)), sz(n)
      {
      if ((!p) && (n!=0))
38
        throw bad_alloc();
Martin Reinecke's avatar
Martin Reinecke committed
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
      }
    ~arr() { dealloc(p); }

    void resize(size_t n)
      {
      if (n==sz) return;
      dealloc(p);
      p = ralloc(n);
      sz = n;
      }

    T &operator[](size_t idx) { return p[idx]; }
    const T &operator[](size_t idx) const { return p[idx]; }

    T *data() { return p; }
    const T *data() const { return p; }

    size_t size() const { return sz; }
  };

class sincos_2pibyn
  {
  private:
    arr<double> data;

    // adapted from https://stackoverflow.com/questions/42792939/
    // CAUTION: this function only works for arguments in the range [-0.25; 0.25]!
    void my_sincosm1pi (double a, double *restrict res)
      {
      double s = a * a;
      /* Approximate cos(pi*x)-1 for x in [-0.25,0.25] */
      double r =     -1.0369917389758117e-4;
      r = fma (r, s,  1.9294935641298806e-3);
      r = fma (r, s, -2.5806887942825395e-2);
      r = fma (r, s,  2.3533063028328211e-1);
      r = fma (r, s, -1.3352627688538006e+0);
      r = fma (r, s,  4.0587121264167623e+0);
      r = fma (r, s, -4.9348022005446790e+0);
      double c = r*s;
      /* Approximate sin(pi*x) for x in [-0.25,0.25] */
      r =             4.6151442520157035e-4;
      r = fma (r, s, -7.3700183130883555e-3);
      r = fma (r, s,  8.2145868949323936e-2);
      r = fma (r, s, -5.9926452893214921e-1);
      r = fma (r, s,  2.5501640398732688e+0);
      r = fma (r, s, -5.1677127800499516e+0);
      s = s * a;
      r = r * s;
      s = fma (a, 3.1415926535897931e+0, r);
      res[0] = c;
      res[1] = s;
      }

    NOINLINE void calc_first_octant(size_t den, double * restrict res)
      {
      size_t n = (den+4)>>3;
      if (n==0) return;
      res[0]=1.; res[1]=0.;
      if (n==1) return;
      size_t l1=(size_t)sqrt(n);
      for (size_t i=1; i<l1; ++i)
        my_sincosm1pi((2.*i)/den,&res[2*i]);
      size_t start=l1;
      while(start<n)
        {
        double cs[2];
        my_sincosm1pi((2.*start)/den,cs);
        res[2*start] = cs[0]+1.;
        res[2*start+1] = cs[1];
        size_t end = l1;
        if (start+end>n) end = n-start;
        for (size_t i=1; i<end; ++i)
          {
          double csx[2]={res[2*i], res[2*i+1]};
          res[2*(start+i)] = ((cs[0]*csx[0] - cs[1]*csx[1] + cs[0]) + csx[0]) + 1.;
          res[2*(start+i)+1] = (cs[0]*csx[1] + cs[1]*csx[0]) + cs[1] + csx[1];
          }
        start += l1;
        }
      for (size_t i=1; i<l1; ++i)
        res[2*i] += 1.;
      }

    NOINLINE void calc_first_quadrant(size_t n, double * restrict res)
      {
      double * restrict p = res+n;
      calc_first_octant(n<<1, p);
      size_t ndone=(n+2)>>2;
      size_t i=0, idx1=0, idx2=2*ndone-2;
      for (; i+1<ndone; i+=2, idx1+=2, idx2-=2)
        {
        res[idx1]   = p[2*i];
        res[idx1+1] = p[2*i+1];
        res[idx2]   = p[2*i+3];
        res[idx2+1] = p[2*i+2];
        }
      if (i!=ndone)
        {
        res[idx1  ] = p[2*i];
        res[idx1+1] = p[2*i+1];
        }
      }

    NOINLINE void calc_first_half(size_t n, double * restrict res)
      {
      int ndone=(n+1)>>1;
      double * p = res+n-1;
      calc_first_octant(n<<2, p);
      int i4=0, in=n, i=0;
      for (; i4<=in-i4; ++i, i4+=4) // octant 0
        {
        res[2*i] = p[2*i4]; res[2*i+1] = p[2*i4+1];
        }
      for (; i4-in <= 0; ++i, i4+=4) // octant 1
        {
        int xm = in-i4;
        res[2*i] = p[2*xm+1]; res[2*i+1] = p[2*xm];
        }
      for (; i4<=3*in-i4; ++i, i4+=4) // octant 2
        {
        int xm = i4-in;
        res[2*i] = -p[2*xm+1]; res[2*i+1] = p[2*xm];
        }
      for (; i<ndone; ++i, i4+=4) // octant 3
        {
        int xm = 2*in-i4;
        res[2*i] = -p[2*xm]; res[2*i+1] = p[2*xm+1];
        }
      }

    NOINLINE void fill_first_quadrant(size_t n, double * restrict res)
      {
      const double hsqt2 = 0.707106781186547524400844362104849;
      size_t quart = n>>2;
      if ((n&7)==0)
        res[quart] = res[quart+1] = hsqt2;
      for (size_t i=2, j=2*quart-2; i<quart; i+=2, j-=2)
        {
        res[j  ] = res[i+1];
        res[j+1] = res[i  ];
        }
      }

    NOINLINE void fill_first_half(size_t n, double * restrict res)
      {
      size_t half = n>>1;
      if ((n&3)==0)
        for (size_t i=0; i<half; i+=2)
          {
          res[i+half]   = -res[i+1];
          res[i+half+1] =  res[i  ];
          }
      else
        for (size_t i=2, j=2*half-2; i<half; i+=2, j-=2)
          {
          res[j  ] = -res[i  ];
          res[j+1] =  res[i+1];
          }
      }

    NOINLINE void fill_second_half(size_t n, double * restrict res)
      {
      if ((n&1)==0)
        for (size_t i=0; i<n; ++i)
          res[i+n] = -res[i];
      else
        for (size_t i=2, j=2*n-2; i<n; i+=2, j-=2)
          {
          res[j  ] =  res[i  ];
          res[j+1] = -res[i+1];
          }
      }

    NOINLINE void sincos_2pibyn_half(size_t n, double * restrict res)
      {
      if ((n&3)==0)
        {
        calc_first_octant(n, res);
        fill_first_quadrant(n, res);
        fill_first_half(n, res);
        }
      else if ((n&1)==0)
        {
        calc_first_quadrant(n, res);
        fill_first_half(n, res);
        }
      else
        calc_first_half(n, res);
      }

  public:
    sincos_2pibyn(size_t n)
      : data(2*n)
      {
      sincos_2pibyn_half(n, data.data());
      fill_second_half(n, data.data());
      }

    double operator[](size_t idx) const { return data[idx]; }
  };

NOINLINE size_t largest_prime_factor (size_t n)
  {
  size_t res=1;
  while ((n&1)==0)
    { res=2; n>>=1; }

  size_t limit=size_t(sqrt(n+0.01));
  for (size_t x=3; x<=limit; x+=2)
  while ((n/x)*x==n)
    {
    res=x;
    n/=x;
    limit=size_t(sqrt(n+0.01));
    }
  if (n>1) res=n;

  return res;
  }

NOINLINE double cost_guess (size_t n)
  {
  const double lfp=1.1; // penalty for non-hardcoded larger factors
  size_t ni=n;
  double result=0.;
  while ((n&1)==0)
    { result+=2; n>>=1; }

  size_t limit=size_t(sqrt(n+0.01));
  for (size_t x=3; x<=limit; x+=2)
  while ((n/x)*x==n)
    {
    result+= (x<=5) ? x : lfp*x; // penalize larger prime factors
    n/=x;
    limit=size_t(sqrt(n+0.01));
    }
  if (n>1) result+=(n<=5) ? n : lfp*n;

  return result*ni;
  }

/* returns the smallest composite of 2, 3, 5, 7 and 11 which is >= n */
NOINLINE size_t good_size(size_t n)
  {
  if (n<=12) return n;

  size_t bestfac=2*n;
  for (size_t f2=1; f2<bestfac; f2*=2)
    for (size_t f23=f2; f23<bestfac; f23*=3)
      for (size_t f235=f23; f235<bestfac; f235*=5)
        for (size_t f2357=f235; f2357<bestfac; f2357*=7)
          for (size_t f235711=f2357; f235711<bestfac; f235711*=11)
            if (f235711>=n) bestfac=f235711;
  return bestfac;
  }

template<typename T> struct cmplx {
  T r, i;
  cmplx() {}
  cmplx(T r_, T i_) : r(r_), i(i_) {}
  cmplx &operator+= (const cmplx &other)
    { r+=other.r; i+=other.i; return *this; }
  template<typename T2>cmplx &operator*= (T2 other)
    { r*=other; i*=other; return *this; }
  cmplx operator+ (const cmplx &other) const
    { return cmplx(r+other.r, i+other.i); }
  cmplx operator- (const cmplx &other) const
    { return cmplx(r-other.r, i-other.i); }
  template<typename T2> auto operator* (const T2 &other) const
    -> cmplx<decltype(r*other)>
    {
    return {r*other, i*other};
    }
  template<typename T2> auto operator* (const cmplx<T2> &other) const
    -> cmplx<decltype(r+other.r)>
    {
    return {r*other.r-i*other.i, r*other.i + i*other.r};
    }
  template<bool bwd, typename T2> auto special_mul (const cmplx<T2> &other) const
    -> cmplx<decltype(r+other.r)>
    {
    if (bwd)
      return {r*other.r-i*other.i, r*other.i + i*other.r};
    else
      return {r*other.r+i*other.i, i*other.r - r*other.i};
    }
};
template<typename T> void PMC(cmplx<T> &a, cmplx<T> &b,
  const cmplx<T> &c, const cmplx<T> &d)
  { a = c+d; b = c-d; }
template<typename T> cmplx<T> conj(const cmplx<T> &a)
  { return {a.r, -a.i}; }

template<typename T> void ROT90(cmplx<T> &a)
  { auto tmp_=a.r; a.r=-a.i; a.i=tmp_; }
template<typename T> void ROTM90(cmplx<T> &a)
  { auto tmp_=-a.r; a.r=a.i; a.i=tmp_; }

#define CH(a,b,c) ch[(a)+ido*((b)+l1*(c))]
#define CC(a,b,c) cc[(a)+ido*((b)+cdim*(c))]
#define WA(x,i) wa[(i)-1+(x)*(ido-1)]

constexpr size_t NFCT=25;

template<typename T0> class cfftp
  {
  private:

    struct fctdata
      {
      size_t fct;
      cmplx<T0> *tw, *tws;
      };

    size_t length, nfct;
    arr<cmplx<T0>> mem;
    fctdata fct[NFCT];

    void add_factor(size_t factor)
      {
359
      if (nfct>=NFCT) throw runtime_error("too many prime factors");
Martin Reinecke's avatar
Martin Reinecke committed
360
361
362
      fct[nfct++].fct = factor;
      }

Martin Reinecke's avatar
Martin Reinecke committed
363
364
template<bool bwd, typename T> NOINLINE void pass2 (size_t ido, size_t l1,
  const T * restrict cc, T * restrict ch, const cmplx<T0> * restrict wa)
Martin Reinecke's avatar
Martin Reinecke committed
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
  {
  constexpr size_t cdim=2;

  if (ido==1)
    for (size_t k=0; k<l1; ++k)
      {
      CH(0,k,0) = CC(0,0,k)+CC(0,1,k);
      CH(0,k,1) = CC(0,0,k)-CC(0,1,k);
      }
  else
    for (size_t k=0; k<l1; ++k)
      {
      CH(0,k,0) = CC(0,0,k)+CC(0,1,k);
      CH(0,k,1) = CC(0,0,k)-CC(0,1,k);
      for (size_t i=1; i<ido; ++i)
        {
        CH(i,k,0) = CC(i,0,k)+CC(i,1,k);
        CH(i,k,1) = (CC(i,0,k)-CC(i,1,k)).template special_mul<bwd>(WA(0,i));
        }
      }
  }

#define PREP3(idx) \
        T t0 = CC(idx,0,k), t1, t2; \
        PMC (t1,t2,CC(idx,1,k),CC(idx,2,k)); \
        CH(idx,k,0)=t0+t1;
#define PARTSTEP3a(u1,u2,twr,twi) \
        { \
        T ca,cb; \
        ca=t0+t1*twr; \
        cb=t2*twi; ROT90(cb); \
        PMC(CH(0,k,u1),CH(0,k,u2),ca,cb) ;\
        }
#define PARTSTEP3b(u1,u2,twr,twi) \
        { \
        T ca,cb,da,db; \
        ca=t0+t1*twr; \
        cb=t2*twi; ROT90(cb); \
        PMC(da,db,ca,cb); \
        CH(i,k,u1) = da.template special_mul<bwd>(WA(u1-1,i)); \
        CH(i,k,u2) = db.template special_mul<bwd>(WA(u2-1,i)); \
        }
Martin Reinecke's avatar
Martin Reinecke committed
407
template<bool bwd, typename T> NOINLINE void pass3 (size_t ido, size_t l1,
Martin Reinecke's avatar
Martin Reinecke committed
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
  const T * restrict cc, T * restrict ch, const cmplx<T0> * restrict wa)
  {
  constexpr size_t cdim=3;
  constexpr T0 tw1r=-0.5, tw1i= (bwd ? 1.: -1.) * 0.86602540378443864676;

  if (ido==1)
    for (size_t k=0; k<l1; ++k)
      {
      PREP3(0)
      PARTSTEP3a(1,2,tw1r,tw1i)
      }
  else
    for (size_t k=0; k<l1; ++k)
      {
      {
      PREP3(0)
      PARTSTEP3a(1,2,tw1r,tw1i)
      }
      for (size_t i=1; i<ido; ++i)
        {
        PREP3(i)
        PARTSTEP3b(1,2,tw1r,tw1i)
        }
      }
  }

Martin Reinecke's avatar
Martin Reinecke committed
434
template<bool bwd, typename T> NOINLINE void pass4 (size_t ido, size_t l1,
Martin Reinecke's avatar
Martin Reinecke committed
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
  const T * restrict cc, T * restrict ch, const cmplx<T0> * restrict wa)
  {
  constexpr size_t cdim=4;

  if (ido==1)
    for (size_t k=0; k<l1; ++k)
      {
      T t1, t2, t3, t4;
      PMC(t2,t1,CC(0,0,k),CC(0,2,k));
      PMC(t3,t4,CC(0,1,k),CC(0,3,k));
      bwd ? ROT90(t4) : ROTM90(t4);
      PMC(CH(0,k,0),CH(0,k,2),t2,t3);
      PMC(CH(0,k,1),CH(0,k,3),t1,t4);
      }
  else
    for (size_t k=0; k<l1; ++k)
      {
      {
      T t1, t2, t3, t4;
      PMC(t2,t1,CC(0,0,k),CC(0,2,k));
      PMC(t3,t4,CC(0,1,k),CC(0,3,k));
      bwd ? ROT90(t4) : ROTM90(t4);
      PMC(CH(0,k,0),CH(0,k,2),t2,t3);
      PMC(CH(0,k,1),CH(0,k,3),t1,t4);
      }
      for (size_t i=1; i<ido; ++i)
        {
        T c2, c3, c4, t1, t2, t3, t4;
        T cc0=CC(i,0,k), cc1=CC(i,1,k),cc2=CC(i,2,k),cc3=CC(i,3,k);
        PMC(t2,t1,cc0,cc2);
        PMC(t3,t4,cc1,cc3);
        bwd ? ROT90(t4) : ROTM90(t4);
        cmplx<T0> wa0=WA(0,i), wa1=WA(1,i),wa2=WA(2,i);
        PMC(CH(i,k,0),c3,t2,t3);
        PMC(c2,c4,t1,t4);
        CH(i,k,1) = c2.template special_mul<bwd>(wa0);
        CH(i,k,2) = c3.template special_mul<bwd>(wa1);
        CH(i,k,3) = c4.template special_mul<bwd>(wa2);
        }
      }
  }

#define PREP5(idx) \
        T t0 = CC(idx,0,k), t1, t2, t3, t4; \
        PMC (t1,t4,CC(idx,1,k),CC(idx,4,k)); \
        PMC (t2,t3,CC(idx,2,k),CC(idx,3,k)); \
        CH(idx,k,0).r=t0.r+t1.r+t2.r; \
        CH(idx,k,0).i=t0.i+t1.i+t2.i;

#define PARTSTEP5a(u1,u2,twar,twbr,twai,twbi) \
        { \
        T ca,cb; \
        ca.r=t0.r+twar*t1.r+twbr*t2.r; \
        ca.i=t0.i+twar*t1.i+twbr*t2.i; \
        cb.i=twai*t4.r twbi*t3.r; \
        cb.r=-(twai*t4.i twbi*t3.i); \
        PMC(CH(0,k,u1),CH(0,k,u2),ca,cb); \
        }

#define PARTSTEP5b(u1,u2,twar,twbr,twai,twbi) \
        { \
        T ca,cb,da,db; \
        ca.r=t0.r+twar*t1.r+twbr*t2.r; \
        ca.i=t0.i+twar*t1.i+twbr*t2.i; \
        cb.i=twai*t4.r twbi*t3.r; \
        cb.r=-(twai*t4.i twbi*t3.i); \
        PMC(da,db,ca,cb); \
        CH(i,k,u1) = da.template special_mul<bwd>(WA(u1-1,i)); \
        CH(i,k,u2) = db.template special_mul<bwd>(WA(u2-1,i)); \
        }
Martin Reinecke's avatar
Martin Reinecke committed
505
template<bool bwd, typename T> NOINLINE void pass5 (size_t ido, size_t l1,
Martin Reinecke's avatar
Martin Reinecke committed
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
  const T * restrict cc, T * restrict ch, const cmplx<T0> * restrict wa)
  {
  constexpr size_t cdim=5;
  constexpr T0 tw1r= 0.3090169943749474241,
               tw1i= (bwd ? 1.: -1.) * 0.95105651629515357212,
               tw2r= -0.8090169943749474241,
               tw2i= (bwd ? 1.: -1.) * 0.58778525229247312917;

  if (ido==1)
    for (size_t k=0; k<l1; ++k)
      {
      PREP5(0)
      PARTSTEP5a(1,4,tw1r,tw2r,+tw1i,+tw2i)
      PARTSTEP5a(2,3,tw2r,tw1r,+tw2i,-tw1i)
      }
  else
    for (size_t k=0; k<l1; ++k)
      {
      {
      PREP5(0)
      PARTSTEP5a(1,4,tw1r,tw2r,+tw1i,+tw2i)
      PARTSTEP5a(2,3,tw2r,tw1r,+tw2i,-tw1i)
      }
      for (size_t i=1; i<ido; ++i)
        {
        PREP5(i)
        PARTSTEP5b(1,4,tw1r,tw2r,+tw1i,+tw2i)
        PARTSTEP5b(2,3,tw2r,tw1r,+tw2i,-tw1i)
        }
      }
  }

#define PREP7(idx) \
        T t1 = CC(idx,0,k), t2, t3, t4, t5, t6, t7; \
        PMC (t2,t7,CC(idx,1,k),CC(idx,6,k)); \
        PMC (t3,t6,CC(idx,2,k),CC(idx,5,k)); \
        PMC (t4,t5,CC(idx,3,k),CC(idx,4,k)); \
        CH(idx,k,0).r=t1.r+t2.r+t3.r+t4.r; \
        CH(idx,k,0).i=t1.i+t2.i+t3.i+t4.i;

#define PARTSTEP7a0(u1,u2,x1,x2,x3,y1,y2,y3,out1,out2) \
        { \
        T ca,cb; \
        ca.r=t1.r+x1*t2.r+x2*t3.r+x3*t4.r; \
        ca.i=t1.i+x1*t2.i+x2*t3.i+x3*t4.i; \
        cb.i=y1*t7.r y2*t6.r y3*t5.r; \
        cb.r=-(y1*t7.i y2*t6.i y3*t5.i); \
        PMC(out1,out2,ca,cb); \
        }
#define PARTSTEP7a(u1,u2,x1,x2,x3,y1,y2,y3) \
        PARTSTEP7a0(u1,u2,x1,x2,x3,y1,y2,y3,CH(0,k,u1),CH(0,k,u2))
#define PARTSTEP7(u1,u2,x1,x2,x3,y1,y2,y3) \
        { \
        T da,db; \
        PARTSTEP7a0(u1,u2,x1,x2,x3,y1,y2,y3,da,db) \
        CH(i,k,u1) = da.template special_mul<bwd>(WA(u1-1,i)); \
        CH(i,k,u2) = db.template special_mul<bwd>(WA(u2-1,i)); \
        }

Martin Reinecke's avatar
Martin Reinecke committed
565
template<bool bwd, typename T> NOINLINE void pass7(size_t ido, size_t l1,
Martin Reinecke's avatar
Martin Reinecke committed
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
  const T * restrict cc, T * restrict ch, const cmplx<T0> * restrict wa)
  {
  constexpr size_t cdim=7;
  constexpr T0 tw1r= 0.623489801858733530525,
               tw1i= (bwd ? 1. : -1.) * 0.7818314824680298087084,
               tw2r= -0.222520933956314404289,
               tw2i= (bwd ? 1. : -1.) * 0.9749279121818236070181,
               tw3r= -0.9009688679024191262361,
               tw3i= (bwd ? 1. : -1.) * 0.4338837391175581204758;

  if (ido==1)
    for (size_t k=0; k<l1; ++k)
      {
      PREP7(0)
      PARTSTEP7a(1,6,tw1r,tw2r,tw3r,+tw1i,+tw2i,+tw3i)
      PARTSTEP7a(2,5,tw2r,tw3r,tw1r,+tw2i,-tw3i,-tw1i)
      PARTSTEP7a(3,4,tw3r,tw1r,tw2r,+tw3i,-tw1i,+tw2i)
      }
  else
    for (size_t k=0; k<l1; ++k)
      {
      {
      PREP7(0)
      PARTSTEP7a(1,6,tw1r,tw2r,tw3r,+tw1i,+tw2i,+tw3i)
      PARTSTEP7a(2,5,tw2r,tw3r,tw1r,+tw2i,-tw3i,-tw1i)
      PARTSTEP7a(3,4,tw3r,tw1r,tw2r,+tw3i,-tw1i,+tw2i)
      }
      for (size_t i=1; i<ido; ++i)
        {
        PREP7(i)
        PARTSTEP7(1,6,tw1r,tw2r,tw3r,+tw1i,+tw2i,+tw3i)
        PARTSTEP7(2,5,tw2r,tw3r,tw1r,+tw2i,-tw3i,-tw1i)
        PARTSTEP7(3,4,tw3r,tw1r,tw2r,+tw3i,-tw1i,+tw2i)
        }
      }
  }

#define PREP11(idx) \
        T t1 = CC(idx,0,k), t2, t3, t4, t5, t6, t7, t8, t9, t10, t11; \
        PMC (t2,t11,CC(idx,1,k),CC(idx,10,k)); \
        PMC (t3,t10,CC(idx,2,k),CC(idx, 9,k)); \
        PMC (t4,t9 ,CC(idx,3,k),CC(idx, 8,k)); \
        PMC (t5,t8 ,CC(idx,4,k),CC(idx, 7,k)); \
        PMC (t6,t7 ,CC(idx,5,k),CC(idx, 6,k)); \
        CH(idx,k,0).r=t1.r+t2.r+t3.r+t4.r+t5.r+t6.r; \
        CH(idx,k,0).i=t1.i+t2.i+t3.i+t4.i+t5.i+t6.i;

#define PARTSTEP11a0(u1,u2,x1,x2,x3,x4,x5,y1,y2,y3,y4,y5,out1,out2) \
        { \
        T ca = t1 + t2*x1 + t3*x2 + t4*x3 + t5*x4 +t6*x5, \
          cb; \
        cb.i=y1*t11.r y2*t10.r y3*t9.r y4*t8.r y5*t7.r; \
        cb.r=-(y1*t11.i y2*t10.i y3*t9.i y4*t8.i y5*t7.i ); \
        PMC(out1,out2,ca,cb); \
        }
#define PARTSTEP11a(u1,u2,x1,x2,x3,x4,x5,y1,y2,y3,y4,y5) \
        PARTSTEP11a0(u1,u2,x1,x2,x3,x4,x5,y1,y2,y3,y4,y5,CH(0,k,u1),CH(0,k,u2))
#define PARTSTEP11(u1,u2,x1,x2,x3,x4,x5,y1,y2,y3,y4,y5) \
        { \
        T da,db; \
        PARTSTEP11a0(u1,u2,x1,x2,x3,x4,x5,y1,y2,y3,y4,y5,da,db) \
        CH(i,k,u1) = da.template special_mul<bwd>(WA(u1-1,i)); \
        CH(i,k,u2) = db.template special_mul<bwd>(WA(u2-1,i)); \
        }

Martin Reinecke's avatar
Martin Reinecke committed
631
template<bool bwd, typename T> NOINLINE void pass11 (size_t ido, size_t l1,
Martin Reinecke's avatar
Martin Reinecke committed
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
  const T * restrict cc, T * restrict ch, const cmplx<T0> * restrict wa)
  {
  const size_t cdim=11;
  const T0 tw1r =        0.8412535328311811688618,
           tw1i = (bwd ? 1. : -1.) * 0.5406408174555975821076,
           tw2r =        0.4154150130018864255293,
           tw2i = (bwd ? 1. : -1.) * 0.9096319953545183714117,
           tw3r =       -0.1423148382732851404438,
           tw3i = (bwd ? 1. : -1.) * 0.9898214418809327323761,
           tw4r =       -0.6548607339452850640569,
           tw4i = (bwd ? 1. : -1.) * 0.755749574354258283774,
           tw5r =       -0.9594929736144973898904,
           tw5i = (bwd ? 1. : -1.) * 0.2817325568414296977114;

  if (ido==1)
    for (size_t k=0; k<l1; ++k)
      {
      PREP11(0)
      PARTSTEP11a(1,10,tw1r,tw2r,tw3r,tw4r,tw5r,+tw1i,+tw2i,+tw3i,+tw4i,+tw5i)
      PARTSTEP11a(2, 9,tw2r,tw4r,tw5r,tw3r,tw1r,+tw2i,+tw4i,-tw5i,-tw3i,-tw1i)
      PARTSTEP11a(3, 8,tw3r,tw5r,tw2r,tw1r,tw4r,+tw3i,-tw5i,-tw2i,+tw1i,+tw4i)
      PARTSTEP11a(4, 7,tw4r,tw3r,tw1r,tw5r,tw2r,+tw4i,-tw3i,+tw1i,+tw5i,-tw2i)
      PARTSTEP11a(5, 6,tw5r,tw1r,tw4r,tw2r,tw3r,+tw5i,-tw1i,+tw4i,-tw2i,+tw3i)
      }
  else
    for (size_t k=0; k<l1; ++k)
      {
      {
      PREP11(0)
      PARTSTEP11a(1,10,tw1r,tw2r,tw3r,tw4r,tw5r,+tw1i,+tw2i,+tw3i,+tw4i,+tw5i)
      PARTSTEP11a(2, 9,tw2r,tw4r,tw5r,tw3r,tw1r,+tw2i,+tw4i,-tw5i,-tw3i,-tw1i)
      PARTSTEP11a(3, 8,tw3r,tw5r,tw2r,tw1r,tw4r,+tw3i,-tw5i,-tw2i,+tw1i,+tw4i)
      PARTSTEP11a(4, 7,tw4r,tw3r,tw1r,tw5r,tw2r,+tw4i,-tw3i,+tw1i,+tw5i,-tw2i)
      PARTSTEP11a(5, 6,tw5r,tw1r,tw4r,tw2r,tw3r,+tw5i,-tw1i,+tw4i,-tw2i,+tw3i)
      }
      for (size_t i=1; i<ido; ++i)
        {
        PREP11(i)
        PARTSTEP11(1,10,tw1r,tw2r,tw3r,tw4r,tw5r,+tw1i,+tw2i,+tw3i,+tw4i,+tw5i)
        PARTSTEP11(2, 9,tw2r,tw4r,tw5r,tw3r,tw1r,+tw2i,+tw4i,-tw5i,-tw3i,-tw1i)
        PARTSTEP11(3, 8,tw3r,tw5r,tw2r,tw1r,tw4r,+tw3i,-tw5i,-tw2i,+tw1i,+tw4i)
        PARTSTEP11(4, 7,tw4r,tw3r,tw1r,tw5r,tw2r,+tw4i,-tw3i,+tw1i,+tw5i,-tw2i)
        PARTSTEP11(5, 6,tw5r,tw1r,tw4r,tw2r,tw3r,+tw5i,-tw1i,+tw4i,-tw2i,+tw3i)
        }
      }
  }

#define CX(a,b,c) cc[(a)+ido*((b)+l1*(c))]
#define CX2(a,b) cc[(a)+idl1*(b)]
#define CH2(a,b) ch[(a)+idl1*(b)]

Martin Reinecke's avatar
Martin Reinecke committed
683
template<bool bwd, typename T> NOINLINE void passg (size_t ido, size_t ip,
Martin Reinecke's avatar
Martin Reinecke committed
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
  size_t l1, T * restrict cc, T * restrict ch, const cmplx<T0> * restrict wa,
  const cmplx<T0> * restrict csarr)
  {
  const size_t cdim=ip;
  size_t ipph = (ip+1)/2;
  size_t idl1 = ido*l1;

  arr<cmplx<T0>> wal(ip);
  wal[0] = cmplx<T0>(1., 0.);
  for (size_t i=1; i<ip; ++i)
    wal[i]=cmplx<T0>(csarr[i].r,bwd ? csarr[i].i : -csarr[i].i);

  for (size_t k=0; k<l1; ++k)
    for (size_t i=0; i<ido; ++i)
      CH(i,k,0) = CC(i,0,k);
  for (size_t j=1, jc=ip-1; j<ipph; ++j, --jc)
    for (size_t k=0; k<l1; ++k)
      for (size_t i=0; i<ido; ++i)
        PMC(CH(i,k,j),CH(i,k,jc),CC(i,j,k),CC(i,jc,k));
  for (size_t k=0; k<l1; ++k)
    for (size_t i=0; i<ido; ++i)
      {
      T tmp = CH(i,k,0);
      for (size_t j=1; j<ipph; ++j)
        tmp+=CH(i,k,j);
      CX(i,k,0) = tmp;
      }
  for (size_t l=1, lc=ip-1; l<ipph; ++l, --lc)
    {
    // j=0
    for (size_t ik=0; ik<idl1; ++ik)
      {
      CX2(ik,l).r = CH2(ik,0).r+wal[l].r*CH2(ik,1).r+wal[2*l].r*CH2(ik,2).r;
      CX2(ik,l).i = CH2(ik,0).i+wal[l].r*CH2(ik,1).i+wal[2*l].r*CH2(ik,2).i;
      CX2(ik,lc).r=-wal[l].i*CH2(ik,ip-1).i-wal[2*l].i*CH2(ik,ip-2).i;
      CX2(ik,lc).i=wal[l].i*CH2(ik,ip-1).r+wal[2*l].i*CH2(ik,ip-2).r;
      }

    size_t iwal=2*l;
    size_t j=3, jc=ip-3;
    for (; j<ipph-1; j+=2, jc-=2)
      {
      iwal+=l; if (iwal>ip) iwal-=ip;
      cmplx<T0> xwal=wal[iwal];
      iwal+=l; if (iwal>ip) iwal-=ip;
      cmplx<T0> xwal2=wal[iwal];
      for (size_t ik=0; ik<idl1; ++ik)
        {
        CX2(ik,l).r += CH2(ik,j).r*xwal.r+CH2(ik,j+1).r*xwal2.r;
        CX2(ik,l).i += CH2(ik,j).i*xwal.r+CH2(ik,j+1).i*xwal2.r;
        CX2(ik,lc).r -= CH2(ik,jc).i*xwal.i+CH2(ik,jc-1).i*xwal2.i;
        CX2(ik,lc).i += CH2(ik,jc).r*xwal.i+CH2(ik,jc-1).r*xwal2.i;
        }
      }
    for (; j<ipph; ++j, --jc)
      {
      iwal+=l; if (iwal>ip) iwal-=ip;
      cmplx<T0> xwal=wal[iwal];
      for (size_t ik=0; ik<idl1; ++ik)
        {
        CX2(ik,l).r += CH2(ik,j).r*xwal.r;
        CX2(ik,l).i += CH2(ik,j).i*xwal.r;
        CX2(ik,lc).r -= CH2(ik,jc).i*xwal.i;
        CX2(ik,lc).i += CH2(ik,jc).r*xwal.i;
        }
      }
    }

  // shuffling and twiddling
  if (ido==1)
    for (size_t j=1, jc=ip-1; j<ipph; ++j, --jc)
      for (size_t ik=0; ik<idl1; ++ik)
        {
        T t1=CX2(ik,j), t2=CX2(ik,jc);
        PMC(CX2(ik,j),CX2(ik,jc),t1,t2);
        }
  else
    {
    for (size_t j=1, jc=ip-1; j<ipph; ++j,--jc)
      for (size_t k=0; k<l1; ++k)
        {
        T t1=CX(0,k,j), t2=CX(0,k,jc);
        PMC(CX(0,k,j),CX(0,k,jc),t1,t2);
        for (size_t i=1; i<ido; ++i)
          {
          T x1, x2;
          PMC(x1,x2,CX(i,k,j),CX(i,k,jc));
          size_t idij=(j-1)*(ido-1)+i-1;
          CX(i,k,j) = x1.template special_mul<bwd>(wa[idij]);
          idij=(jc-1)*(ido-1)+i-1;
          CX(i,k,jc) = x2.template special_mul<bwd>(wa[idij]);
          }
        }
    }
  }

#undef CH2
#undef CX2
#undef CX

Martin Reinecke's avatar
Martin Reinecke committed
784
template<bool bwd, typename T> NOINLINE void pass_all(T c[], T0 fact)
Martin Reinecke's avatar
Martin Reinecke committed
785
786
787
788
789
790
791
792
793
794
795
796
  {
  if (length==1) return;
  size_t l1=1;
  arr<T> ch(length);
  T *p1=c, *p2=ch.data();

  for(size_t k1=0; k1<nfct; k1++)
    {
    size_t ip=fct[k1].fct;
    size_t l2=ip*l1;
    size_t ido = length/l2;
    if     (ip==4)
Martin Reinecke's avatar
Martin Reinecke committed
797
      pass4<bwd> (ido, l1, p1, p2, fct[k1].tw);
Martin Reinecke's avatar
Martin Reinecke committed
798
    else if(ip==2)
Martin Reinecke's avatar
Martin Reinecke committed
799
      pass2<bwd>(ido, l1, p1, p2, fct[k1].tw);
Martin Reinecke's avatar
Martin Reinecke committed
800
    else if(ip==3)
Martin Reinecke's avatar
Martin Reinecke committed
801
      pass3<bwd> (ido, l1, p1, p2, fct[k1].tw);
Martin Reinecke's avatar
Martin Reinecke committed
802
    else if(ip==5)
Martin Reinecke's avatar
Martin Reinecke committed
803
      pass5<bwd> (ido, l1, p1, p2, fct[k1].tw);
Martin Reinecke's avatar
Martin Reinecke committed
804
    else if(ip==7)
Martin Reinecke's avatar
Martin Reinecke committed
805
      pass7<bwd> (ido, l1, p1, p2, fct[k1].tw);
Martin Reinecke's avatar
Martin Reinecke committed
806
    else if(ip==11)
Martin Reinecke's avatar
Martin Reinecke committed
807
      pass11<bwd> (ido, l1, p1, p2, fct[k1].tw);
Martin Reinecke's avatar
Martin Reinecke committed
808
809
    else
      {
Martin Reinecke's avatar
Martin Reinecke committed
810
      passg<bwd>(ido, ip, l1, p1, p2, fct[k1].tw, fct[k1].tws);
Martin Reinecke's avatar
Martin Reinecke committed
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
      swap(p1,p2);
      }
    swap(p1,p2);
    l1=l2;
    }
  if (p1!=c)
    {
    if (fact!=1.)
      for (size_t i=0; i<length; ++i)
        c[i] = ch[i]*fact;
    else
      memcpy (c,p1,length*sizeof(T));
    }
  else
    if (fact!=1.)
      for (size_t i=0; i<length; ++i)
        c[i] *= fact;
  }

#undef WA
#undef CC
#undef CH
#undef PMC

  public:
    template<typename T> NOINLINE void forward(T c[], T0 fct)
Martin Reinecke's avatar
Martin Reinecke committed
837
      { pass_all<false>(c, fct); }
Martin Reinecke's avatar
Martin Reinecke committed
838
839

    template<typename T> NOINLINE void backward(T c[], T0 fct)
Martin Reinecke's avatar
Martin Reinecke committed
840
      { pass_all<true>(c, fct); }
Martin Reinecke's avatar
Martin Reinecke committed
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917

  private:
    NOINLINE void factorize()
      {
      nfct=0;
      size_t len=length;
      while ((len&3)==0)
        { add_factor(4); len>>=2; }
      if ((len&1)==0)
        {
        len>>=1;
        // factor 2 should be at the front of the factor list
        add_factor(2);
        swap(fct[0].fct, fct[nfct-1].fct);
        }
      size_t maxl=(size_t)(sqrt((double)len))+1;
      for (size_t divisor=3; (len>1)&&(divisor<maxl); divisor+=2)
        if ((len%divisor)==0)
          {
          while ((len%divisor)==0)
            {
            add_factor(divisor);
            len/=divisor;
            }
          maxl=size_t(sqrt((double)len))+1;
          }
      if (len>1) add_factor(len);
      }

    NOINLINE size_t twsize() const
      {
      size_t twsize=0, l1=1;
      for (size_t k=0; k<nfct; ++k)
        {
        size_t ip=fct[k].fct, ido= length/(l1*ip);
        twsize+=(ip-1)*(ido-1);
        if (ip>11)
          twsize+=ip;
        l1*=ip;
        }
      return twsize;
      }

    NOINLINE void comp_twiddle()
      {
      sincos_2pibyn twid(length);
      size_t l1=1;
      size_t memofs=0;
      for (size_t k=0; k<nfct; ++k)
        {
        size_t ip=fct[k].fct, ido=length/(l1*ip);
        fct[k].tw=mem.data()+memofs;
        memofs+=(ip-1)*(ido-1);
        for (size_t j=1; j<ip; ++j)
          for (size_t i=1; i<ido; ++i)
            {
            fct[k].tw[(j-1)*(ido-1)+i-1].r = twid[2*j*l1*i];
            fct[k].tw[(j-1)*(ido-1)+i-1].i = twid[2*j*l1*i+1];
            }
        if (ip>11)
          {
          fct[k].tws=mem.data()+memofs;
          memofs+=ip;
          for (size_t j=0; j<ip; ++j)
            {
            fct[k].tws[j].r = twid[2*j*l1*ido];
            fct[k].tws[j].i = twid[2*j*l1*ido+1];
            }
          }
        l1*=ip;
        }
      }

  public:
    NOINLINE cfftp(size_t length_)
      : length(length_), nfct(0)
      {
918
      if (length==0) throw runtime_error("zero length FFT requested");
Martin Reinecke's avatar
Martin Reinecke committed
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
      if (length==1) return;
      factorize();
      mem.resize(twsize());
      comp_twiddle();
      }
  };

template<typename T0> class fftblue
  {
  private:
    size_t n, n2;
    cfftp<T0> plan;
    arr<T0> mem;
    T0 *bk, *bkf;

    template<bool bwd, typename T> NOINLINE
    void fft(T c[], T0 fct)
      {
      constexpr int isign = bwd ? 1 : -1;
      arr<T> akf(2*n2);

      /* initialize a_k and FFT it */
      for (size_t m=0; m<2*n; m+=2)
        {
        akf[m]   = c[m]*bk[m]   -isign*c[m+1]*bk[m+1];
        akf[m+1] = isign*c[m]*bk[m+1] + c[m+1]*bk[m];
        }
      for (size_t m=2*n; m<2*n2; ++m)
        akf[m]=0.*c[0];

      plan.forward ((cmplx<T> *)akf.data(),1.);

      /* do the convolution */
      for (size_t m=0; m<2*n2; m+=2)
        {
        T im = -isign*akf[m]*bkf[m+1] + akf[m+1]*bkf[m];
        akf[m  ]  =  akf[m]*bkf[m]   + isign*akf[m+1]*bkf[m+1];
        akf[m+1]  = im;
        }

      /* inverse FFT */
      plan.backward ((cmplx<T> *)akf.data(),1.);

      /* multiply by b_k */
      for (size_t m=0; m<2*n; m+=2)
        {
        c[m]   = fct*(bk[m]  *akf[m] - isign*bk[m+1]*akf[m+1]);
        c[m+1] = fct*(isign*bk[m+1]*akf[m] + bk[m]  *akf[m+1]);
        }
      }

  public:
    NOINLINE fftblue(size_t length)
      : n(length), n2(good_size(n*2-1)), plan(n2), mem(2*(n+n2)),
        bk(mem.data()), bkf(mem.data()+2*n)
      {
      /* initialize b_k */
      sincos_2pibyn tmp(2*n);
      bk[0] = 1;
      bk[1] = 0;

      size_t coeff=0;
      for (size_t m=1; m<n; ++m)
        {
        coeff+=2*m-1;
        if (coeff>=2*n) coeff-=2*n;
        bk[2*m  ] = tmp[2*coeff  ];
        bk[2*m+1] = tmp[2*coeff+1];
        }

      /* initialize the zero-padded, Fourier transformed b_k. Add normalisation. */
      T0 xn2 = 1./n2;
      bkf[0] = bk[0]*xn2;
      bkf[1] = bk[1]*xn2;
      for (size_t m=2; m<2*n; m+=2)
        {
        bkf[m]   = bkf[2*n2-m]   = bk[m]   *xn2;
        bkf[m+1] = bkf[2*n2-m+1] = bk[m+1] *xn2;
        }
      for (size_t m=2*n;m<=(2*n2-2*n+1);++m)
        bkf[m]=0.;
      plan.forward((cmplx<T0> *)bkf,1.);
      }

    template<typename T> void backward(T c[], T0 fct)
      { fft<true>(c,fct); }

    template<typename T> void forward(T c[], T0 fct)
      { fft<false>(c,fct); }
    };

template<typename T0> class pocketfft_c
  {
  private:
    unique_ptr<cfftp<T0>> packplan;
    unique_ptr<fftblue<T0>> blueplan;
    size_t len;

  public:
    NOINLINE pocketfft_c(size_t length)
      : packplan(nullptr), blueplan(nullptr), len(length)
      {
1021
      if (length==0) throw runtime_error("zero-length FFT requested");
Martin Reinecke's avatar
Martin Reinecke committed
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
      if ((length<50) || (largest_prime_factor(length)<=sqrt(length)))
        {
        packplan=unique_ptr<cfftp<T0>>(new cfftp<T0>(length));
        return;
        }
      double comp1 = cost_guess(length);
      double comp2 = 2*cost_guess(good_size(2*length-1));
      comp2*=1.5; /* fudge factor that appears to give good overall performance */
      if (comp2<comp1) // use Bluestein
        blueplan=unique_ptr<fftblue<T0>>(new fftblue<T0>(length));
      else
        packplan=unique_ptr<cfftp<T0>>(new cfftp<T0>(length));
      }

    template<typename T> void backward(T c[], T0 fct)
      {
      packplan ? packplan->backward((cmplx<T> *)c,fct)
               : blueplan->backward(c,fct);
      }

    template<typename T> void forward(T c[], T0 fct)
      {
      packplan ? packplan->forward((cmplx<T> *)c,fct)
               : blueplan->forward(c,fct);
      }

    size_t length() const { return len; }
  };

Martin Reinecke's avatar
Martin Reinecke committed
1051
struct diminfo
Martin Reinecke's avatar
Martin Reinecke committed
1052
  { size_t n; int64_t s; };
Martin Reinecke's avatar
Martin Reinecke committed
1053
1054
1055
1056
1057
1058
class multiarr
  {
  private:
    vector<diminfo> dim;

  public:
Martin Reinecke's avatar
Martin Reinecke committed
1059
    multiarr (size_t ndim_, const size_t *n, const int64_t *s)
Martin Reinecke's avatar
Martin Reinecke committed
1060
1061
1062
1063
1064
1065
1066
      {
      dim.reserve(ndim_);
      for (size_t i=0; i<ndim_; ++i)
        dim.push_back({n[i], s[i]});
      }
    size_t ndim() const { return dim.size(); }
    size_t size(size_t i) const { return dim[i].n; }
Martin Reinecke's avatar
Martin Reinecke committed
1067
    int64_t stride(size_t i) const { return dim[i].s; }
Martin Reinecke's avatar
Martin Reinecke committed
1068
1069
1070
1071
1072
1073
1074
  };

class multi_iter
  {
  private:
    vector<diminfo> dim;
    vector<size_t> pos;
Martin Reinecke's avatar
Martin Reinecke committed
1075
1076
    size_t ofs_, len;
    int64_t str;
Martin Reinecke's avatar
Martin Reinecke committed
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
    bool done_;

  public:
    multi_iter(const multiarr &arr, size_t idim)
      : pos(arr.ndim()-1, 0), ofs_(0), len(arr.size(idim)),
        str(arr.stride(idim)), done_(false)
      {
      dim.reserve(arr.ndim()-1);
      for (size_t i=0; i<arr.ndim(); ++i)
        if (i!=idim)
          {
          dim.push_back({arr.size(i), arr.stride(i)});
          done_ = done_ || (arr.size(i)==0);
          }
      }
    void advance()
      {
      for (int i=pos.size()-1; i>=0; --i)
        {
        ++pos[i];
        ofs_ += dim[i].s;
        if (pos[i] < dim[i].n)
          return;
        pos[i] = 0;
        ofs_ -= dim[i].n*dim[i].s;
        }
      done_ = true;
      }
    bool done() const { return done_; }
    size_t offset() const { return ofs_; }
    size_t length() const { return len; }
Martin Reinecke's avatar
Martin Reinecke committed
1108
    int64_t stride() const { return str; }
Martin Reinecke's avatar
Martin Reinecke committed
1109
1110
1111
  };

template<typename T> void pocketfft_general_c(int ndim, const size_t *shape,
Martin Reinecke's avatar
Martin Reinecke committed
1112
  const int64_t *stride_in, const int64_t *stride_out, int nax,
Martin Reinecke's avatar
Martin Reinecke committed
1113
1114
1115
1116
1117
1118
1119
1120
  const size_t *axes, bool forward, const cmplx<T> *data_in,
  cmplx<T> *data_out, T fct)
  {
  // allocate temporary 1D array storage, if necessary
  size_t tmpsize = 0;
  for (int iax=0; iax<nax; ++iax)
    {
    bool inplace = false;
Martin Reinecke's avatar
Martin Reinecke committed
1121
    int64_t stride = (iax==0) ? stride_in[axes[iax]] : stride_out[axes[iax]];
Martin Reinecke's avatar
Martin Reinecke committed
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
    if (stride==1)
      if ((iax>0) || (data_in==data_out))
        inplace = true;
    if (!inplace)
      if (shape[axes[iax]] > tmpsize) tmpsize = shape[axes[iax]];
    }
  arr<cmplx<T>> tdata(tmpsize);
  multiarr a_in(ndim, shape, stride_in), a_out(ndim, shape, stride_out);
  unique_ptr<pocketfft_c<T>> plan;

  for (int iax=0; iax<nax; ++iax)
    {
    int axis = axes[iax];
    multi_iter it_in(a_in, axis), it_out(a_out, axis);
    bool inplace = (data_in==data_out) && (it_in.stride()==1);
    if ((!plan) || (it_in.length()!=plan->length()))
      plan.reset(new pocketfft_c<T>(it_in.length()));
    while (!it_in.done())
      {
      if (inplace)
        {
        cmplx<T> *d = data_out+it_in.offset();
        forward ? plan->forward((T *)d, fct)
                : plan->backward((T *)d, fct);
        }
      else
        {
        for (size_t i=0; i<it_in.length(); ++i)
          tdata[i] = data_in[it_in.offset()+i*it_in.stride()];
        forward ? plan->forward((T *)tdata.data(), fct)
                : plan->backward((T *)tdata.data(), fct);
        for (size_t i=0; i<it_out.length(); ++i)
          data_out[it_out.offset()+i*it_out.stride()] = tdata[i];
        }
      it_in.advance();
      if (!inplace)
        it_out.advance();
      }
    // after the first dimension, take data from output array
    a_in = a_out;
    data_in = data_out;
    // factor has been applied, use 1 for remaining axes
Martin Reinecke's avatar
Martin Reinecke committed
1164
    fct = T(1);
Martin Reinecke's avatar
Martin Reinecke committed
1165
1166
1167
1168
1169
    }
  }

namespace py = pybind11;

Martin Reinecke's avatar
Martin Reinecke committed
1170
1171
auto c64 = py::dtype("complex64");
auto c128 = py::dtype("complex128");
Martin Reinecke's avatar
Martin Reinecke committed
1172
1173
1174
1175

py::array execute(const py::array &in, bool fwd)
  {
  py::array res;
Martin Reinecke's avatar
Martin Reinecke committed
1176
1177
  vector<size_t> dims(in.ndim()), axes(in.ndim());
  vector<int64_t> s_i(in.ndim()), s_o(in.ndim());
Martin Reinecke's avatar
Martin Reinecke committed
1178
1179
  for (int i=0; i<in.ndim(); ++i)
    dims[i] = in.shape(i);
Martin Reinecke's avatar
Martin Reinecke committed
1180
  if (in.dtype().is(c128))
Martin Reinecke's avatar
Martin Reinecke committed
1181
    res = py::array_t<complex<double>>(dims);
Martin Reinecke's avatar
Martin Reinecke committed
1182
  else if (in.dtype().is(c64))
Martin Reinecke's avatar
Martin Reinecke committed
1183
    res = py::array_t<complex<float>>(dims);
Martin Reinecke's avatar
Martin Reinecke committed
1184
1185
1186
1187
  else throw runtime_error("unsupported data type");
  for (int i=0; i<in.ndim(); ++i)
    {
    s_i[i] = in.strides(i)/in.itemsize();
Martin Reinecke's avatar
Martin Reinecke committed
1188
1189
    if (s_i[i]*in.itemsize() != in.strides(i))
      throw runtime_error("weird strides");
Martin Reinecke's avatar
Martin Reinecke committed
1190
    s_o[i] = res.strides(i)/res.itemsize();
Martin Reinecke's avatar
Martin Reinecke committed
1191
1192
    if (s_o[i]*res.itemsize() != res.strides(i))
      throw runtime_error("weird strides");
Martin Reinecke's avatar
Martin Reinecke committed
1193
1194
    axes[i] = i;
    }
Martin Reinecke's avatar
Martin Reinecke committed
1195
  if (in.dtype().is(c128))
Martin Reinecke's avatar
Martin Reinecke committed
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
    pocketfft_general_c<double>(in.ndim(), dims.data(),
    s_i.data(), s_o.data(), in.ndim(),
    axes.data(), fwd, (const cmplx<double> *)in.data(),
    (cmplx<double> *)res.mutable_data(), 1.);
  else
    pocketfft_general_c<float>(in.ndim(), dims.data(),
    s_i.data(), s_o.data(), in.ndim(),
    axes.data(), fwd, (const cmplx<float> *)in.data(),
    (cmplx<float> *)res.mutable_data(), 1.);
  return res;
  }
py::array fftn(const py::array &in)
  { return execute(in, true); }
py::array ifftn(const py::array &in)
  { return execute(in, false); }
} // unnamed namespace

PYBIND11_MODULE(pypocketfft, m)
  {
  m.def("fftn",&fftn);
  m.def("ifftn",&ifftn);
  }